Anthropic published a study yesterday that should make every developer building multilingual AI products stop and think. After analyzing 309,815 real conversations across three Claude models and 20 languages, the researchers found that Claude's expressed values — warmth, rigor, deference, caution — shift systematically depending on what language you speak to it in.

Ask Claude to evaluate a business plan in Hindi, and you'll get encouragement, polite framing, and affirmation. Ask in Russian, and you'll get precision, skepticism, and a demand for evidence. Same model. Same weights. Different values.

This isn't a bug report. It's a feature of how training data works. But it raises a question Anthropic itself can't yet answer: if the model's values shift with language, which version is the real Claude?

The Four Axes

The study builds on Anthropic's earlier "Values in the Wild" research, which identified 3,307 distinct value terms in Claude's responses. That's too many to compare systematically, so the research team used dimensionality reduction to compress them into four core axes:

Axis	One Side	The Other
Deference vs. Caution	Accommodation, engagement, respecting user preferences	Risk mitigation, responsible communication, pushing back
Warmth vs. Rigor	Encouragement, positivity, comfort, humor	Accuracy, transparency, precision, evidence-seeking
Depth vs. Brevity	Nuance, critical thinking, substance	Short answers, getting to the point
Candor vs. Execution	Intellectual honesty, admitting limitations	Results orientation, getting things done

These four axes capture about 15% of the behavioral variation — meaning 85% of what makes Claude different across conversations isn't explained by this framework. That's a significant caveat the paper emphasizes, and one worth keeping in mind.

The Model Personalities

One of the more striking findings is how cleanly the axes recover the informal personality profiles developers have been noticing for months:

Sonnet 4.6 is the Affirmer. It leans deference (+0.14σ) and warmth (+0.17σ). It affirms user ideas, mirrors tone, uses humor, and offers comfort. Great for creative collaboration, bad for when you need someone to tell you your idea is bad.
Opus 4.6 is the Executor. It's balanced but slightly rigor-leaning (+0.10σ) with a tendency toward brevity (+0.08σ). Stays within scope, gets to the point, doesn't volunteer extra. The "do exactly what I asked" model.
Opus 4.7 is the Critic — and it's not close. It leads on caution (+0.24σ) and depth (+0.23σ) with a comfortable margin. It pushes back on false assumptions, flags risks unprompted, gives candid critiques, and acknowledges its own errors. If you want a rubber duck that actually pushes back, this is your model.

These differences are small in absolute magnitude but structured and consistent. More importantly, they line up with how developers describe these models in practice — which is a healthy validation that the methodology is capturing something real.

The Language Effect — Bigger, and Less Understood

This is where the study gets uncomfortable. The Warmth vs. Rigor axis shows the largest language-driven variation:

Language	Profile
Hindi	Warmest. Politeness, humor, playfulness, affirmations.
Arabic	Warmest tier + highest deference + brevity.
English	Most cautious. Most rigorous. Favoring depth (long-winded).
Russian	Most rigorous tier. Questioning assumptions, demanding evidence.
Dutch	Most candid — particularly open about its own limitations.
Indonesian	Least candid — prefers to execute without expressing doubt.

Anthropic puts it bluntly in the paper: "Two people asking for feedback on the same business plan, one in Hindi and one in Russian, may come away with different impressions of its quality because Claude expressed different values in how it framed its assessment."

That's not a hypothetical. That's a direct consequence of how the training data is distributed.

Why This Happens (And Why They Can't Be Sure)

The paper is refreshingly honest about the limits of its explanation. The researchers identify data imbalance and data composition as likely drivers — training data isn't evenly distributed across languages, and some languages are overrepresented in professional writing (English, Russian) while others skew casual or conversational (Hindi, Arabic).

But here's the key quote: "We don't yet know which properties of our training data drive these differences."

They also flag a methodological concern worth noting: the value labels were assigned by Claude Sonnet 4.6 itself (a model within the family being studied). Anthropic verified via manual review and 800 translated conversations, but acknowledges they "couldn't fully rule out remaining effects." That's self-labeling bias, and it's a significant limitation.

The four axes explaining only 15% of variation should also give us pause. The headline patterns are real, but they're a thin slice of a much more complex picture.

What This Means for Developers

If you're building multilingual AI products, this study has real practical implications:

Language choice is a behavioral tuning parameter. A support bot in Hindi will naturally feel warmer than one in English. An Arabic QA system will be more deferential. A Russian coding assistant will be more rigorous. You need to account for this in your prompt engineering, not assume consistent behavior.

Model choice compounds the effect. Running Sonnet 4.6 in Arabic vs. Opus 4.7 in English isn't just two models and two languages — it's a double multiplier on behavior. Your testing pipeline needs to span both dimensions.

The "real Claude" question matters for safety. If values shift with language, which version's values are correct? The warm, deferential Claude in Hindi? The cautious, rigorous Claude in English? Anthropic frames this as an open question — "whether the language differences represent desirable adaptation to different speech communities or unintended training effects" — but for safety-critical deployments, the answer can't stay open forever.

The Bottom Line

This study is the first systematic attempt to map how a frontier model's values vary across languages, and it should be required reading for any team shipping multilingual AI products. The four axes provide a useful vocabulary for describing model behavior, and the language findings confirm what many suspected but few could prove: your experience of Claude depends heavily on what language you prompt in.

The uncomfortable truth is that Anthropic is publishing this research without being able to fully explain it. That's honest science, but it's also a reminder that alignment — ensuring models behave consistently and predictably — gets exponentially harder when you add language as a variable.

The full study is available on Anthropic's research page.